Deep Learning Applications on Omics Data and Diagnostics

M.Phil. Thesis | Kadir Has University — Computational Biology & Bioinformatics | 2018

This repository contains the Jupyter Notebook implementations from my master's thesis, Deep Learning Applications on Biological Data. The work applies Softmax Regression, Feed Forward Neural Networks (FFNN), and Long Short-Term Memory (LSTM) networks to four biological and medical datasets, demonstrating how increasing model complexity improves classification accuracy.

Stack: Python · TensorFlow · Keras · NumPy · Pandas · Scikit-learn · Matplotlib

---

Notebooks

Softmax Classification

Notebook	Description
Softmax_Multiclass_Classification_4Datasets.ipynb	Softmax multiclass classification on all four datasets (Anuran Call, Thyroid, E. coli, HIV) with min-max input normalization and loss/accuracy plotting

Softmax with Keras — Bias-Variance Tradeoff & Optimizer Comparison

Notebook	Description
Softmax_Keras_BiasVariance_Optimizer_Comparison.ipynb	Softmax on four datasets implemented in Keras; optimizer comparison (ADAM, SGD, RMSprop); bias-variance tradeoff analysis; introduces LSTM

Feed Forward Neural Networks & LSTM (Keras)

Notebook	Description
Anuran_Frog_Species_Softmax_Baseline.ipynb	Softmax baseline for Anuran Call (frog species) classification — 15 classes, 22 MFCC features
Anuran_Frog_Species_FFNN_Regularization.ipynb	FFNN model selection and regularization (L2, Dropout) for frog species — accuracy improved from 78% to 95%
Ecoli_Protein_Localization_Softmax.ipynb	Softmax classification for E. coli protein subcellular localization — 336 instances, 7 attributes, 8 classes
HIV_Cleavage_FFNN_Model_Selection.ipynb	FFNN model selection for HIV-1 protease cleavage site prediction — binary classification on 6,590 octamer sequences
HIV_Cleavage_FFNN_Regularization.ipynb	Regularization tuning (L2 and Dropout) on the best HIV FFNN model; final test set evaluation

---

Datasets

Anuran Call — 7,195 syllables from 60 audio recordings of 15 frog species; 22 MFCCs per syllable.

Thyroid Patients — 72,000 patient records with 21 clinical attributes; task: classify as healthy, hyperthyroid, or hypothyroid.

E. coli Protein Localization — 336 instances, 7 biophysical attributes; task: predict protein subcellular localization site (8 classes).

HIV Cleavage Sites — 6,590 amino acid octamer sequences (concatenated from four datasets); task: binary classification of HIV-1 protease cleavage sites.

---

Results

Dataset	Softmax	FFNN
Anuran (Frog Species)	78%	95%
Thyroid Patients	high	—
E. coli Localization	moderate	—
HIV Cleavage Sites	81%	improved with LSTM

---

How to Run

pip install tensorflow keras numpy pandas scikit-learn matplotlib jupyter
jupyter notebook

---

Background

Completed as part of the M.Phil. program in Computational Biology and Bioinformatics at Kadir Has University, Istanbul, under the supervision of Assoc. Prof. Cem Özen.